基于DirectML的Windows OCR推理方案-平芜编程栈

1. 引言

2. 优势

3. 效果

4. OCR 模型选择与准备

5. WinForm 界面实现

6. 下载

1. 引言

本方案旨在利用 DirectML 的硬件加速能力，在 Windows 系统上部署和运行一个轻量级、高精度的 OCR 模型。

2. 优势

其核心优势在于：

跨硬件兼容：DirectML 支持集成显卡、独立显卡（NVIDIA/AMD/Intel）以及 CPU，确保在不同 Windows 设备上都能获得加速。
低延迟推理：通过 DirectX 12 底层接口直接与硬件通信，减少中间层开销，实现端到端的高效推理。
易于集成：提供 C++ 和 WinRT API，可方便地集成到桌面应用、UWP 应用或服务中。
模型格式支持：支持 ONNX 模型格式，便于使用主流深度学习框架（如 PyTorch, TensorFlow）训练并导出的模型。

3. 效果

4. OCR 模型选择与准备

PaddleOCR 轻量级模型：百度开源的 OCR 工具包，提供了优秀的轻量级中英文识别模型。

5. WinForm 界面实现

为了将 OCR 推理功能集成到 Windows 桌面应用中，我们可以使用 WinForm 创建一个简单的图形界面。以下是一个基本的 WinForm 窗体实现，包含图像选择、OCR 识别和结果显示功能。

namespace OnnxRuntimeForm { public partial class Form1 : Form { private IntPtr engine = IntPtr.Zero; private Bitmap currentImage = null; private string imgPath = null; private List<OcrResult> ltOCRResult = new List<OcrResult>(); private StringBuilder OCRResultInfo = new StringBuilder(); private StringBuilder OCRResultAllInfo = new StringBuilder(); private Pen pen = new Pen(Brushes.Red, 2f); private const string BaseModelDir = @"inference"; public class OcrResult { public List<List<double>> box { get; set; } public double score { get; set; } public string text { get; set; } } [DllImport("OnnxRuntimeDll.dll", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Unicode)] static extern int init(out IntPtr engine, [MarshalAs(UnmanagedType.U1)] bool use_gpu, int gpu_id, string det_model_dir, int limit_side_len, double det_db_thresh, double det_db_box_thresh, double det_db_unclip_ratio, [MarshalAs(UnmanagedType.U1)] bool use_dilation, [MarshalAs(UnmanagedType.U1)] bool cls, [MarshalAs(UnmanagedType.U1)] bool use_angle_cls, string cls_model_dir, double cls_thresh, double cls_batch_num, string rec_model_dir, string rec_char_dict_path, int rec_batch_num, int rec_img_h, int rec_img_w, int predictor_num, StringBuilder msg); [DllImport("OnnxRuntimeDll.dll", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Unicode)] static extern int ocr(IntPtr engine, int rows, int cols, int channels, IntPtr data, StringBuilder msg, out IntPtr ocr_result, out int ocr_result_len); [DllImport("OnnxRuntimeDll.dll", CallingConvention = CallingConvention.Cdecl, CharSet = CharSet.Unicode)] static extern int destroy(IntPtr engine, StringBuilder msg); public Form1() { InitializeComponent(); } private void Form1_Load(object sender, EventArgs e) { rdoV6Tiny.Checked = true; chkGPU.Checked = false; chkCls.Checked = false; UpdateModelPaths(); txtUnclipRatio.Text = "1.8"; LoadModel(); } private void rdoModel_CheckedChanged(object sender, EventArgs e) { if (((RadioButton)sender).Checked) { UpdateModelPaths(); } } private void UpdateModelPaths() { string det = ""; string rec = ""; string dict = ""; bool v6 = false; if (rdoV5Mobile.Checked) { det = Path.Combine(BaseModelDir, "PP-OCRv5_mobile_det_onnx.onnx"); rec = Path.Combine(BaseModelDir, "PP-OCRv5_mobile_rec_onnx.onnx"); dict = Path.Combine(BaseModelDir, "ppocrv5_dict.txt"); } else if (rdoV5Server.Checked) { det = Path.Combine(BaseModelDir, "PP-OCRv5_server_det_infer.onnx"); rec = Path.Combine(BaseModelDir, "PP-OCRv5_server_rec_infer.onnx"); dict = Path.Combine(BaseModelDir, "ppocrv5_dict.txt"); } else if (rdoV6Tiny.Checked) { det = Path.Combine(BaseModelDir, "PP-OCRv6_tiny_det.onnx"); rec = Path.Combine(BaseModelDir, "PP-OCRv6_tiny_rec.onnx"); dict = Path.Combine(BaseModelDir, "PP-OCRv6_tiny_rec_dict.txt"); v6 = true; } else if (rdoV6Small.Checked) { det = Path.Combine(BaseModelDir, "PP-OCRv6_small_det.onnx"); rec = Path.Combine(BaseModelDir, "PP-OCRv6_small_rec.onnx"); dict = Path.Combine(BaseModelDir, "PP-OCRv6_small_rec_dict.txt"); v6 = true; } else if (rdoV6Medium.Checked) { det = Path.Combine(BaseModelDir, "PP-OCRv6_medium_det.onnx"); rec = Path.Combine(BaseModelDir, "PP-OCRv6_medium_rec.onnx"); dict = Path.Combine(BaseModelDir, "PP-OCRv6_medium_rec_dict.txt"); v6 = true; } txtDetModel.Text = det; txtRecModel.Text = rec; txtDict.Text = dict; if (v6) { chkCls.Checked = false; chkCls.Enabled = false; } else { chkCls.Enabled = true; } } private void btnInit_Click(object sender, EventArgs e) { richTextBox1.Clear(); if (engine != IntPtr.Zero) { UnloadModel(); } LoadModel(); } private void btnSelect_Click(object sender, EventArgs e) { using (OpenFileDialog ofd = new OpenFileDialog()) { ofd.Filter = "图片|*.bmp;*.jpg;*.jpeg;*.tiff;*.tif;*.png;*.gif|所有文件|*.*"; if (ofd.ShowDialog() == DialogResult.OK) { imgPath = ofd.FileName; currentImage?.Dispose(); currentImage = new Bitmap(imgPath); pictureBox1.Image = currentImage; richTextBoxResult.Clear(); } } } private void btnOCR_Click(object sender, EventArgs e) { if (engine == IntPtr.Zero) { MessageBox.Show("请先初始化模型！"); return; } if (imgPath == null || currentImage == null) { MessageBox.Show("请先选择图片！"); return; } btnSelect.Enabled = false; btnOCR.Enabled = false; richTextBoxResult.Clear(); OCRResultInfo.Clear(); OCRResultAllInfo.Clear(); var msgTemp = new StringBuilder(1024); var stopwatch = System.Diagnostics.Stopwatch.StartNew(); IntPtr strPtr = IntPtr.Zero; int ocr_result_len = 0; string ocr_result = string.Empty; int res; byte[] bgrData; using (Bitmap bgrBitmap = LoadBgrBitmap(imgPath, out bgrData)) { GCHandle handle = GCHandle.Alloc(bgrData, GCHandleType.Pinned); try { res = ocr2(engine, bgrBitmap.Height, bgrBitmap.Width, 3, handle.AddrOfPinnedObject(), msgTemp, out strPtr, out ocr_result_len); } finally { handle.Free(); } } if (strPtr != IntPtr.Zero && ocr_result_len > 0) { byte[] buffer = new byte[ocr_result_len]; Marshal.Copy(strPtr, buffer, 0, ocr_result_len); ocr_result = Encoding.UTF8.GetString(buffer); Marshal.FreeCoTaskMem(strPtr); strPtr = IntPtr.Zero; } stopwatch.Stop(); double totalTime = stopwatch.Elapsed.TotalMilliseconds; OCRResultAllInfo.AppendLine($"耗时: {totalTime:F2}ms"); OCRResultAllInfo.AppendLine("---------------------------"); OCRResultInfo.AppendLine($"耗时: {totalTime:F2}ms"); OCRResultInfo.AppendLine("---------------------------"); if (res == 0) { ltOCRResult = ParseOCRResults(ocr_result); ltOCRResult.Sort((a, b) => { double ax1 = a.box[0][0], ay1 = a.box[0][1], ax2 = a.box[0][0], ay2 = a.box[0][1]; double bx1 = b.box[0][0], by1 = b.box[0][1], bx2 = b.box[0][0], by2 = b.box[0][1]; foreach (var p in a.box) { ax1 = Math.Min(ax1, p[0]); ay1 = Math.Min(ay1, p[1]); ax2 = Math.Max(ax2, p[0]); ay2 = Math.Max(ay2, p[1]); } foreach (var p in b.box) { bx1 = Math.Min(bx1, p[0]); by1 = Math.Min(by1, p[1]); bx2 = Math.Max(bx2, p[0]); by2 = Math.Max(by2, p[1]); } double y_overlap = Math.Max(0, Math.Min(ay2, by2) - Math.Max(ay1, by1)); double h = Math.Min(ay2 - ay1, by2 - by1); if (h > 0 && y_overlap / h > 0.5) return ax1.CompareTo(bx1); return ay1.CompareTo(by1); }); OCRResultAllInfo.Append(FormatJson(ocr_result)); if (ltOCRResult.Count == 0) { OCRResultInfo.AppendLine("[未识别到文字]"); OCRResultAllInfo.AppendLine("[未识别到文字]"); } using (Graphics graphics = Graphics.FromImage(currentImage)) { foreach (OcrResult item in ltOCRResult) { OCRResultInfo.AppendLine(item.text); if (item.box != null && item.box.Count >= 4) { Point[] points = item.box.Select(p => new Point((int)p[0], (int)p[1])).ToArray(); graphics.DrawPolygon(pen, points); } } } pictureBox1.Image = null; pictureBox1.Image = currentImage; UpdateRichText(); } else { if (strPtr != IntPtr.Zero) Marshal.FreeCoTaskMem(strPtr); MessageBox.Show("识别失败：" + msgTemp.ToString()); } btnSelect.Enabled = true; btnOCR.Enabled = true; } private void chkAllInfo_CheckedChanged(object sender, EventArgs e) { UpdateRichText(); } private void UpdateRichText() { richTextBoxResult.Text = chkAllInfo.Checked ? OCRResultAllInfo.ToString() : OCRResultInfo.ToString(); } private double ParseDouble(TextBox tb, double defaultValue) { if (double.TryParse(tb.Text.Trim(), out double v)) return v; return defaultValue; } private int ParseInt(TextBox tb, int defaultValue) { if (int.TryParse(tb.Text.Trim(), out int v)) return v; return defaultValue; } private void LoadModel() { var msgTemp = new StringBuilder(1024); bool use_gpu = chkGPU.Checked; int gpu_id = 0; string det_model_dir = txtDetModel.Text.Trim(); int limit_side_len = ParseInt(txtLimitSideLen, 960); double det_db_thresh = ParseDouble(txtDbThresh, 0.3); double det_db_box_thresh = ParseDouble(txtBoxThresh, 0.6); double det_db_unclip_ratio = ParseDouble(txtUnclipRatio, 1.6); bool use_dilation = true; bool use_angle_cls = chkCls.Checked && chkCls.Enabled; bool cls = use_angle_cls; string cls_model_dir = use_angle_cls ? Path.Combine(BaseModelDir, "PP-OCRv5_mobile_cls_onnx.onnx") : ""; double cls_thresh = 0.9; double cls_batch_num = 1; string rec_model_dir = txtRecModel.Text.Trim(); string rec_char_dict_path = txtDict.Text.Trim(); int rec_batch_num = ParseInt(txtRecBatchNum, 8); int rec_img_h = 48; int rec_img_w = 960; int predictor_num = 4; AppendStatus("正在初始化模型..."); AppendStatus("det: " + det_model_dir); AppendStatus("rec: " + rec_model_dir); AppendStatus("dict: " + rec_char_dict_path); AppendStatus("device: " + (use_gpu ? "GPU" : "CPU")); AppendStatus($"db_thresh={det_db_thresh}, box_thresh={det_db_box_thresh}, unclip={det_db_unclip_ratio}"); int res = init(out engine, use_gpu, gpu_id, det_model_dir, limit_side_len, det_db_thresh, det_db_box_thresh, det_db_unclip_ratio, use_dilation, cls, use_angle_cls, cls_model_dir, cls_thresh, cls_batch_num, rec_model_dir, rec_char_dict_path, rec_batch_num, rec_img_h, rec_img_w, predictor_num, msgTemp); if (res == 0) { AppendStatus("模型加载成功: " + msgTemp.ToString()); } else { AppendStatus("模型加载失败: " + msgTemp.ToString()); MessageBox.Show("模型加载失败：" + msgTemp.ToString()); engine = IntPtr.Zero; } } private void UnloadModel() { if (engine != IntPtr.Zero) { var msgTemp = new StringBuilder(1024); destroy(engine, msgTemp); AppendStatus("释放成功: " + msgTemp.ToString()); engine = IntPtr.Zero; } } private void AppendStatus(string text) { richTextBox1.AppendText($"[{DateTime.Now:HH:mm:ss}] {text}{Environment.NewLine}"); } private Bitmap LoadBgrBitmap(string path, out byte[] bgrData) { using (Bitmap source = new Bitmap(path)) { Bitmap bitmap = new Bitmap(source.Width, source.Height, PixelFormat.Format24bppRgb); using (Graphics g = Graphics.FromImage(bitmap)) { g.DrawImage(source, 0, 0, source.Width, source.Height); } Rectangle rect = new Rectangle(0, 0, bitmap.Width, bitmap.Height); BitmapData data = bitmap.LockBits(rect, ImageLockMode.ReadOnly, PixelFormat.Format24bppRgb); try { int rowBytes = bitmap.Width * 3; bgrData = new byte[rowBytes * bitmap.Height]; for (int y = 0; y < bitmap.Height; y++) { IntPtr src = IntPtr.Add(data.Scan0, y * data.Stride); Marshal.Copy(src, bgrData, y * rowBytes, rowBytes); } } finally { bitmap.UnlockBits(data); } return bitmap; } } private List<OcrResult> ParseOCRResults(string json) { try { var serializer = new JavaScriptSerializer(); return serializer.Deserialize<List<OcrResult>>(json) ?? new List<OcrResult>(); } catch (Exception ex) { MessageBox.Show("JSON 解析失败：" + ex.Message); return new List<OcrResult>(); } } private string FormatJson(string json) { try { var serializer = new JavaScriptSerializer(); var obj = serializer.Deserialize<object>(json); return serializer.Serialize(obj); } catch { return json; } } private void Form1_FormClosing(object sender, FormClosingEventArgs e) { UnloadModel(); currentImage?.Dispose(); pen?.Dispose(); } } }